Bregman Conditional Random Fields: Sequence Labeling with Parallelizable Inference Algorithms

TL;DR

This paper introduces Bregman conditional random fields (BCRF), a new sequence labeling model that enables fast parallel inference and learning, matching or surpassing traditional CRFs in speed and accuracy.

Contribution

The paper presents BCRF, a novel discriminative model that allows parallelizable inference algorithms and extends learning methods to partial labels.

Findings

BCRF achieves comparable accuracy to CRFs with faster inference.

BCRF outperforms mean field methods in constrained settings.

The approach enables efficient learning from partial labels.

Abstract

We propose a novel discriminative model for sequence labeling called Bregman conditional random fields (BCRF). Contrary to standard linear-chain conditional random fields, BCRF allows fast parallelizable inference algorithms based on iterative Bregman projections. We show how such models can be learned using Fenchel-Young losses, including extension for learning from partial labels. Experimentally, our approach delivers comparable results to CRF while being faster, and achieves better results in highly constrained settings compared to mean field, another parallelizable alternative.